In-depth study is not in line with the rumors, but it still promises

Explanation: We have increased our in-depth study for a long time. The time has come to accept it as an appendix, not a substitute for human intelligence.

Image: Getty Images / iStockphoto / nicescene

A few years ago I said the clause “machine learning will eliminate the need for radiologists.” That was not my wise prediction. However, to my dismay, I was joined by some of the greatest in-depth study experts, such as Jeffrey Hinton, who announced in 2016 that this was “absolutely clear”. [that] five years of in-depth study works better than trained radiologists.

He was wrong. I was wrong. And as a field, we are all mistaken about how fast learning, a field of mechanical engineering, is progressing.

Or not really “progressing” because of in-depth study is go fast. What is it no but the work is progressing to the point where it is distracting. The key to appreciating in-depth learning, writes Gary Marcus, scientist and founder of Geometric Intelligence, a machine learning company that was acquired by Uber in 2016, is to recognize that this model recognition tool is “the best when we everything needed is rough-and-tumble results, in which stocks are low and results are completely voluntary. ”

In other words, when cars can replace people rather than replace them.

The game goes into in-depth study aspects

In-depth study is essentially a way to fit a pattern on a scale. No human can search huge chunks of data to discover examples of this information – machines can. Conversely, machines struggle when presented with external indications that are easy for a person to understand, but contradict the information that machines are trained with. Cars can’t make sense – people can. (Well, most people can … often!)

SEE: The ethics policy of artificial intelligence (TechRepublic Premium)

Jared Kaplan of OpenAI stated that the problem is not about the cause, but on the scale. The more data you enter into machines, the closer the machines become to the repetition of the human mind. This view is incorrect.

You don’t have to take my word for it. Just look around. Select any AI / ML system you want. None of them have come close to copying even the ordinary human mind, as they have fallen short in conveying a real understanding of what they mean. This does not mean that it is useless. Away from it. No, it’s arguing that we need to allow people to be people and cars to be cars and find ways to connect our forces.

The truth about real machine learning

We should also try to find an ML / in-depth study of problems that can be solved more easily with simple math, following Noah Laurent’s considerations (“data scientists basically just do arithmetic”). Or as by Amazon applied scientist Eugene Yan“The first rule of machine learning [is to] start without machine learning. ”

If we make an effort understand data, not just broken numbers, we need to be more careful about how we use machines (i.e., ML / AI) at our disposal. In addition to a quote from Laurent, “Laurent’s approach to data science is correct today, as he said a few years ago:“ There is a very small set of business problems that can be better solved through machine learning; most of them just need good information and understanding. ‘”So, he said, instead of overloading in-depth / ML learning models with expectations, we need to do” SQL queries to get data, … arithmetic the main thing to address in this regard. data (calculation differences, percentages, etc.), results graphs and [writing] clauses or recommendations. ”

SEE: Rental Kit: Artificial Intelligence Architect (TechRepublic)

You know: the kind of work we’ve been doing for decades, long before in-depth study became serious.

Back to Ian. For a successful ML project, “You need data. To support your data flow, you need a strong pipeline. And most of all, you need high-quality labels.” This last point emphasizes the need to familiarize yourself with your data: To show well giving it you need to understand the data to some extent.This should all be done before you start throwing random data into an in-depth learning algorithm and praying for results.

This requires even more symbiosis between humans and machines. None replaces the other. As Mary Shacklett of TechRepublic recently wrote, “Great transfers don’t work in a vacuum. It coordinates with human decision makers and works in a symbiotic mode with people to make or carry out optimal decisions or actions. ”Thus, this will help if we over-sell future in-depth training, machine learning and intelligence. to stop artificiality and instead focus on the need to better integrate human intelligence with brute force and the compatibility of a machine-based model.

Disclosure: I work for MongoDB, but the ideas expressed here are only mine.

Leave a Comment